Dissertation - Open Access
Doctor of Philosophy (PhD)
Geospatial Science and Engineering
Geoffrey M. Henebry
Harmonized Landsat Sentinel-2, Land Cover / Land Use, Landsat Analysis Ready, Land Surface Phenology
Understanding rapid land change in the U.S. NGP region is not only critical for management and conservation of prairie habitats and ecosystem services, but also for projecting production of crops and biofuels and the impacts of land conversion on water quality and rural transportation infrastructure. Hence, it raises the need for an LCLU dataset with good spatiotemporal coverage as well as consistent accuracy through time to enable change analysis. This dissertation aims (1) to develop a novel classification method, which utilizes time series images from comparable sensors, from the perspective of land surface phenology, and (2) to apply the land cover/land use dataset generated from the phenometrically-based classification approach to quantify crop expansion in South Dakota. A novel classification approach from the perspective of land surface phenology (LSP) uses rich time series datasets. First, surface reflectance products at 30 m spatial resolution from Landsat Collection-1, its newer structure—Landsat Analysis Ready Data, and the Harmonized Landsat Sentinel-2 (HLS) data are used to construct vegetation index time series, including the Enhanced Vegetation Index (EVI), and the 2-band EVI (EVI2), and various spectral variables (spectral band and normalized ratio composites). MODIS Level-3 Land Surface Temperature & Emissivity 8-day composite products at 1 km spatial resolution from both the Aqua and Terra satellites are used to compute accumulated growing degree-days (AGDD) time series. The EVI/EVI2 and AGDD time series are then fitted by two different land surface phenology models: the Convex Quadratic model and the Hybrid Piecewise Logistic Model. Suites of phenometrics are derived from the two LSP models and spectral variables and input to Random Forest Classifiers (RFC) to map land cover of sample areas in South Dakota. The results indicate that classifications using only phenometrics can accurately map major crops in the study area but show limited accuracy for non-vegetated land covers. RFC models using the combined spectralphenological variables can achieve higher accuracies than those using either spectral variables or phenometrics alone, especially for the barren/developed class. Among all sampling designs, the “same distribution” models—proportional distribution of the sample is like proportional distribution of the population—tends to yield best land cover prediction. A “same distribution” random sample dataset covering approximately 0.25% or more of the study area appears to achieve an accurate land cover map. To characterize crop expansion in South Dakota, a trajectory-based analysis, which considers the entire land cover dataset generated from the LSP-based classifications, is proposed to improve change detection. An estimated cropland expansion of 5,447 km2 (equivalent to 14% of the existing cropland area) occurred between 2007 and 2015, which matches more closely the reports from the National Agriculture Statistics Service—NASS (5,921 km2) and the National Resources Inventory—NRI (5,034 km2) than an estimation from a bi-temporal change approach (8,018 km2). Cropland gains were mostly concentrated in 10 counties in northern and central South Dakota. An evaluation of land suitability for crops using the Soil Survey Geographic Database—SSURGO indicates a scarcity in high-quality arable land available for cropland expansion.
Number of Pages
South Dakota State University
In Copyright - Educational Use Permitted
Nguyen, Lan Hoang, "Augmenting Land Cover/Land Use Classification by Incorporating Information from Land Surface Phenology: An Application to Quantify Recent Cropland Expansion in South Dakota" (2019). Electronic Theses and Dissertations. 3357.